dialog design 4 speech & natural language

Fall 2002 CS/PSY 6750 1

Dialog Design 4Speech & Natural Language

Agenda• What is speech?• When to use speech• SHW discussion• Speech output• Speech input• Designing the speech interaction

Fall 2002 CS/PSY 6750 2

A Voice Interface

Fall 2002 CS/PSY 6750 3

When to Use Speech

• Hands busy• Mobility required• Eyes occupied• Conditions preclude use of keyboard• Visual impairment• Physical limitation

Fall 2002 CS/PSY 6750 4

Topic Discussion

• Is speech appropriate to this task? Airline info system, telephone based

• Was it well done? Acoustics Technical implementation (recognition,

etc.) Interface flow

• What could have been better?

Fall 2002 CS/PSY 6750 5

Speech

• What is speech? Vibrations of vocal cords creates sound

“ahh” Mouth, throat, tongue, lips shape sound

• English speech 40 phonemes; 24 consonants, 16 vowels

• Sounds transmit “language”

Fall 2002 CS/PSY 6750 6

Waveform & Spectrogram

• Speech does not equal written language

Fall 2002 CS/PSY 6750 7

Parsing Sentences

"I told him to go back where he came from, but he wouldn't listen."

Fall 2002 CS/PSY 6750 8

Speech Input

• Speaker recognition• Speech recognition• Natural language understanding

Fall 2002 CS/PSY 6750 9

Speaker Recognition

• Tell which person it is (voice print)

• Could also be important for monitoring meetings, determining speaker

Fall 2002 CS/PSY 6750 10

Speech Recognition

• Primarily identifying words

• Improving all the time• Commercial systems:

IBM ViaVoice, Dragon Dictate, ...

Fall 2002 CS/PSY 6750 11

Recognition Dimensions

• Speaker dependent/independent Parametric patterns are sensitive to

speaker With training (dependent) can get better

• Vocabulary Some have 50,000+ words

• Isolated word vs. continuous speech Continuous: where words stop & begin Typically a pattern match, no context used

Did youvs.Didja

Fall 2002 CS/PSY 6750 12

Recognition Systems

• Typical system has 5 components: Speech capture device - Has analog -> digital

converter Digital Signal Processor - Gets word boundaries,

scales, filters, cuts out extra stuff Preprocessed signal storage - Processed speech

buffered for recognition algorithm Reference speech patterns - Stored templates or

generative speech models for comparisons Pattern matching algorithm - Goodness of fit

from templates/model to user’s speech

Fall 2002 CS/PSY 6750 13

Errors

• Systems make four types of errors: Substitution - one for another Rejection - detected, but not recognized Insertion - added Deletion - not detected

• Which is more common, dangerous?

Fall 2002 CS/PSY 6750 14

Natural Language Understanding

• Putting meaning to the words

• Input might be speech or could be typed in

• Holy grail of Artificial Intelligence problems

Fall 2002 CS/PSY 6750 15

NL Factors/Terms

• Syntactic Grammar or structure

• Prosodic Inflection, stress, pitch, timing

• Pragmatic Situated context of utterance, location,

time• Semantic

Meaning of words

Fall 2002 CS/PSY 6750 16

SR/NLU Advantages

• Easy to learn and remember• Powerful• Fast, efficient (not always)• Little screen real estate

Fall 2002 CS/PSY 6750 17

SR/NLU Disadvantages

• Doesn’t work good enough yet• Assumes knowledge of problem

domain Not prompted, like menus

• Requires typing skill (if keyboard)• Enhancements are invisible• Expensive to implement

Fall 2002 CS/PSY 6750 18

Recall

• A natural language interface need not be speech

• A speech interface need not use natural language (might be more command language-like)

• Wizard of Oz evaluations are particularly useful in this area

Fall 2002 CS/PSY 6750 19

Speech Output

• Male or female voice? Technical issues (freq. response of phone) User preference (depends on the application)

• Rate of speech Technically up to 550 wpm! Depends on listener (blind: 150-300 wpm)

• Synthesized or Pre-recorded? Synthesized: Better coverage, flexibility Recorded: Better quality, acceptance

Fall 2002 CS/PSY 6750 20

Speech Output

• Synthesis Quality depends on software ($$) Influence of vocabulary and phrase choices

• Recorded segments Store tones, then put them together The transitions are difficult (e.g., numbers)

• Numbers Record three versions (rise, flat, fall) Logic to determine which version to play

What was the airline system’s output like ??

Fall 2002 CS/PSY 6750 21

Designing the Interaction

• Constrain vocabulary Limit valid commands Structure questions wisely (Yes/No) Manage the interaction Examples from the airline systems?

• Slow speech rate, but concise phrases• Design for failsafe error recovery• Process preview & progress indicator

Fall 2002 CS/PSY 6750 22

Speech Tools/Toolkits

• Java Speech SDK FreeTTS 1.1.1 http://

freetts.sourceforge.net/docs/index.php "For 3/4 or 75% of his time, Dr. Walker practices for $90 a visit on Dr. Dr., next to King Philip X

of St. Lameer St. in Nashua NH."

• IBM JavaBeans for speech• Visual/Real Basic speech SDK• OS capabilities (speech recognition

and synthesis built in to OS) (TextEdit)• VoiceXML

Talking Clock